3. Some concepts on experimental design

1 Need for a good control

• A good control group is crucial.

• To assess the effect of an intervention we need to compare a test and control group

• This is often not possible in a pretest/posttest design: e.g. effect before and after administring a drug without the use of a placebo group

• Groups in an observational study are often not comparable: advanced statistical methods are required to draw causal conclusions.

• Double blinding

• We have to be aware of confounding!

• Randomized studies: random assignment of subjects in the study to the different treatment arms \(\rightarrow\) comparable groups.

2 Randomisation

• Randomisation completely at random (no systematic allocation)

2.1 Simple randomisation

• Can lead to differences in the number of experimental units in each treatment arm

• in 5% of the cases we might observe an imbalance of
  • of at least 60:40 in a study with 100 subjects
  • of at least 531:469 in a study with 1000 subjects

• This imbalance is not problematic but causes a loss in precision.

2.2 Balanced randomisation

• Equal numbers of each treatment are assigned to a block of 2 or 4 patients.
  • 1. AB, (2) BA
  • 1. AABB, (2) ABAB, (3) ABBA, (4) BABA, (5) BAAB, (6) BBAA

• Gebalanced randomisation ensures \(\pm\) the same number of people in the control and the treatment arm of the experiment.

• Does not make that we have an equal number of males with and without the treatment, etc.

• In small studies it is possible that the groups are inbalanced in other characteristics (e.g. gender, race, age …)

• This is not problematic because it occurs at random, but, again it causes a loss in precision.

2.3 Stratified randomization**

• The imbalance according to for instance gender can be avoided using stratified randomisation: balanced randomisatie per stratum

Stratified Randomisatie

3 Blocking

3.1 Gene expression example

• dm: diabetic medium, nd: non diabetic medium, co: control

• 4 bio-reps, 2 techreps/biorep

• dm: diabetic medium, nd: non diabetic medium, co: control
• 4 bio-reps, 2 techreps/biorep, 2 plates A & B

• treatment and plate almost entirely confounded

3.2 Nature methods: Points of significance - Blocking

4 Sample size

• The sample size and the design are crucial

• The larger the sample size the more precise the results

5 Wrap-up

• Sample size is very important

• To assess the effect of a treatment we should compare comparable and representative groups of subjects with and without the treatment (a good control!).

• In observational studies the researcher cannot choose the treatment. It was the patient or their MD who had chosen it

• In experimental studies the researcher assigns the treatment

• Confounding can be avoided via randomisation

• We can also correct for confounding in the statistical analysis for the confounders that have been registered.